Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

Multi-agent AI research promises a path to develop human-like and human-compatible intelligent technologies that complement the solipsistic view of other approaches, which mostly do not consider interactions between agents. Aiming to make progress in this direction, the Melting Pot contest 2023 focused on the problem of cooperation among interacting agents and challenged researchers to push the boundaries of multi-agent reinforcement learning (MARL) for mixed-motive games. The contest leveraged the Melting Pot environment suite to rigorously evaluate how well agents can adapt their cooperative skills to interact with novel partners in unforeseen situations. Unlike other reinforcement learning challenges, this challenge focused on social rather than environmental generalization. In particular, a population of agents performs well in Melting Pot when its component individuals are adept at finding ways to cooperate both with others in their population and with strangers. Thus Melting Pot measures cooperative intelligence.The contest attracted over 600 participants across 100+ teams globally and was a success on multiple fronts: (i) it contributed to our goal of pushing the frontiers of MARL towards building more cooperatively intelligent agents, evidenced by several submissions that outperformed established baselines; (ii) it attracted a diverse range of participants, from independent researchers to industry affiliates and academic labs, both with strong background and new interest in the area alike, broadening the field’s demographic and intellectual diversity; and (iii) analyzing the submitted agents provided important insights, highlighting areas for improvement in evaluating agents' cooperative intelligence. This paper summarizes the design aspects and results of the contest and explores the potential of Melting Pot as a benchmark for studying Cooperative AI. We further analyze the top solutions and conclude with a discussion on promising directions for future research.

Existing evaluation suites for multi-agent reinforcement learning (MARL) do not assess generalization to novel situations as their primary objective (unlike supervised learning benchmarks). Our contribution, Melting Pot, is a MARL evaluation suite that fills this gap and uses reinforcement learning to reduce the human labor required to create novel test scenarios. This works because one agent’s behavior constitutes (part of) another agent’s environment. To demonstrate scalability, we have created over 80 unique test scenarios covering a broad range of research topics such as social dilemmas, reciprocity, resource sharing, and task partitioning. We apply these test scenarios to standard MARL training algorithms, and demonstrate how Melting Pot reveals weaknesses not apparent from training performance alone.

This paper investigates generalisation in multi-agent games, where the generality of the agent can be evaluated by playing against opponents it hasn’t seen during training. We propose two new games with concealed information and complex, non-transitive reward structure (think rock-paper-scissors). It turns out that most current deep reinforcement learning methods fail to efficiently explore the strategy space, thus learning policies that generalise poorly to unseen opponents. We then propose a novel hierarchical agent architecture, where the hierarchy is grounded in the game-theoretic structure of the game – the top level chooses strategic responses to opponents, while the low level implements them into policy over primitive actions. This grounding facilitates credit assignment across the levels of hierarchy. Our experiments show that the proposed hierarchical agent is capable of generalisation to unseen opponents, while conventional baselines fail to generalise whatsoever.

We introduce FeUdal Networks (FuNs): a novel architecture for hierarchical reinforcement learning. Our approach is inspired by the feudal reinforcement learning proposal of Dayan and Hinton, and gains power and efficacy by decoupling end-to-end learning across multiple levels – allowing it to utilise different resolutions of time. Our framework employs a Manager module and a Worker module. The Manager operates at a slower time scale and sets abstract goals which are conveyed to and enacted by the Worker. The Worker generates primitive actions at every tick of the environment. The decoupled structure of FuN conveys several benefits – in addition to facilitating very long timescale credit assignment it also encourages the emergence of sub-policies associated with different goals set by the Manager. These properties allow FuN to dramatically outperform a strong baseline agent on tasks that involve long-term credit assignment or memorisation.

We present a novel deep recurrent neural network architecture that learns to build implicit plans in an end-to-end manner purely by interacting with an environment in reinforcement learning setting. The network builds an internal plan, which is continuously updated upon observation of the next input from the environment. It can also partition this internal representation into contiguous sub-sequences by learning for how long the plan can be committed to -- i.e. followed without replaning. Combining these properties, the proposed model, dubbed STRategic Attentive Writer (STRAW) can learn high-level, temporally abstracted macro-actions of varying lengths that are solely learnt from data without any prior information. These macro-actions enable both structured exploration and economic computation. We experimentally demonstrate that STRAW delivers strong improvements on several ATARI games by employing temporally extended planning strategies (e.g. Ms. Pacman and Frostbite). It is at the same time a general algorithm that can be applied on any sequence data. To that end, we also show that when trained on text prediction task, STRAW naturally predicts frequent n-grams (instead of macro-actions), demonstrating the generality of the approach.

We present Context Forest (ConF) — a technique for predicting properties of the objects in an image based on its global appearance. Compared to standard nearest- neighbour techniques, ConF is more accurate, fast and memory efficient. We demonstrate ConF by predicting three properties: aspects of appearance, location in the image, and class membership. In extensive experiments we show that (i) ConF can automatically select which components of a multi-component detector to run on a given test image, obtaining a considerable speed-up for detectors trained from large sets (10 for EE-SVMs [36] and 2 for DPM [21]); (ii) ConF can improve object detection performance by removing false positive detections at unlikely locations (+2% mAP), and by (iii) removing false positives produced by classes unlikely to be present in the image (+5% mAP on a 200-class dataset [2])

We propose a method for annotating the location of objects in ImageNet. Traditionally, this is cast as an image window classification problem, where each window is considered independently and scored based on its appearance alone. Instead, we propose a method which scores each candidate window in the context of all other windows in the image, taking into account their similarity in appearance space as well as their spatial relations in the image plane. We devise a fast and exact procedure to optimize our scoring function over all candidate windows in an image, and we learn its parameters using structured output regression. We demonstrate on 92000 images from ImageNet that this significantly improves localization over recent techniques that score windows in isolation.

Object class detectors typically apply a window classifier to all the windows in a large set, either in a sliding window manner or using object proposals. In this paper, we develop an active search strategy that sequentially chooses the next window to evaluate based on all the information gathered before. This results in a substantial reduction in the number of classifier evaluations and in a more elegant approach in general. Our search strategy is guided by two forces. First, we exploit context as the statistical relation between the appearance of a window and its location relative to the object, as observed in the training set. This enables to jump across distant regions in the image (e.g. observing a sky region suggests that cars might be far below) and is done efficiently in a Random Forest framework. Second, we exploit the score of the classifier to attract the search to promising areas surrounding a highly scored window, and to keep away from areas near low scored ones. Our search strategy can be applied on top of any classifier as it treats it as a black-box. In experiments with R-CNN on the challenging SUN2012 dataset, our method matches the detection accuracy of evaluating all windows independently, while evaluating 9× fewer windows.

We present a method for calibrating the Ensemble of Exemplar SVMs model. Unlike the standard approach, which calibrates each SVM independently, our method optimizes their joint performance as an ensemble. We formulate joint calibration as a constrained optimization problem and devise an efficient optimization algorithm to find its global optimum. The algorithm dynamically discards parts of the solution space that cannot contain the optimum early on, making the optimization computationally feasible. We experiment with EE-SVM trained on state-of-the-art CNN descriptors. Results on the ILSVRC 2014 and PASCAL VOC 2007 datasets show that (i) our joint calibration procedure outperforms independent calibration on the task of classifying windows as belonging to an object class or not; and (ii) this improved window classifier leads to better performance on the object detection task.

We propose a method for knowledge transfer between semantically related classes in ImageNet. By transferring knowledge from the images that have bounding-box annotations to the others, our method is capable of automatically populating ImageNet with many more bounding-boxes. The underlying assumption that objects from semantically related classes look alike is formalized in our novel Associative Embedding (AE) representation. AE recovers the latent low-dimensional space of appearance variations among image windows. The dimensions of AE space tend to correspond to aspects of window appearance (e.g. side view, close up, background). We model the overlap of a window with an object using Gaussian Processes (GP) regression, which spreads annotation smoothly through AE space. The probabilistic nature of GP allows our method to perform self-assessment, i.e. assigning a quality estimate to its own output. It enables trading off the amount of returned annotations for their quality. A large scale experiment on 219 classes and 0.5 million images demonstrates that our method outperforms state-of-the-art methods and baselines for object localization. Using self-assessment we can automatically return bounding-box annotations for 51% of all images with high localization accuracy (i.e. 71% average overlap with ground-truth).

We address the problem of semantic segmentation: classifying each pixel in an image according to the semantic class it belongs to (e.g. dog, road, car). Most existing methods train from fully supervised images, where each pixel is annotated by a class label. To reduce the annotation effort, recently a few weakly supervised approaches emerged. These require only image labels indicating which classes are present. Although their performance reaches a satisfactory level, there is still a substantial gap between the accuracy of fully and weakly supervised methods. We address this gap with a novel active learning method specifically suited for this setting. We model the problem as a pairwise CRF and cast active learning as finding its most informative nodes. These nodes induce the largest expected change in the overall CRF state, after revealing their true label. Our criterion is equivalent to maximizing an upper-bound on accuracy gain. Experiments on two data-sets show that our method achieves 97% percent of the accuracy of the corresponding fully supervised model, while querying less than 17% of the (super-)pixel labels.

We address the problem of weakly supervised semantic segmentation. The training images are labeled only by the classes they contain, not by their location in the image. On test images instead, the method must predict a class label for every pixel. Our goal is to enable segmentation algorithms to use multiple visual cues in this weakly supervised setting, analogous to what is achieved by fully supervised methods. However, it is difficult to assess the relative usefulness of different visual cues from weakly supervised training data. We define a parametric family of structured models, were each model weights visual cues in a different way. We propose a Maximum Expected Agreement model selection principle that evaluates the quality of a model from the family without looking at superpixel labels. Searching for the best model is a hard optimization problem, which has no analytic gradient and multiple local optima. We cast it as a Bayesian optimization problem and propose an algorithm based on Gaussian processes to efficiently solve it. Our second contribution is an Extremely Randomized Hashing Forest that represents diverse superpixel features as a sparse binary vector. It enables using appearance models of visual classes that are fast at training and testing and yet accurate. Experiments on the SIFT-flow dataset show a significant improvement over previous weakly supervised methods and even over some fully supervised methods.

We propose a novel method for weakly supervised semantic segmentation. Training images are labeled only by the classes they contain, not by their location in the image. On test images instead, the method predicts a class label for every pixel. Our main innovation is a multi-image model (MIM) - a graphical model for recovering the pixel labels of the training images. The model connects superpixels from all training images in a data-driven fashion, based on their appearance similarity. For generalizing to new test images we integrate them into MIM using a learned multiple kernel metric, instead of learning conventional classifiers on the recovered pixel labels. We also introduce an “objectness” potential, that helps separating objects (e.g. car, dog, human) from background classes (e.g. grass, sky, road). In experiments on the MSRC 21 dataset and the LabelMe subset of [18], our technique outperforms previous weakly supervised methods and achieves accuracy comparable with fully supervised methods.

We address the task of learning a semantic segmentation from weakly supervised data. Our aim is to devise a system that predicts an object label for each pixel by making use of only image level labels during training - the information whether a certain object is present or not in the image. Such coarse tagging of images is faster and easier to obtain as opposed to the tedious task of pixelwise labeling required in state of the art systems. We cast this task naturally as a multiple instance learning (MIL) problem. We use Semantic Texton Forest (STF) as the basic framework and extend it for the MIL setting. We make use of multitask learning (MTL) to regularize our solution. Here, an external task of geometric context estimation is used to improve on the task of semantic segmentation. We report experimental results on the MSRC21 and the very challenging VOC2007 datasets. On MSRC21 dataset we are able, by using 276 weakly labeled images, to achieve the performance of a supervised STF trained on pixelwise labeled training set of 56 images, which is a significant reduction in supervision needed.